Home Technology peripherals AI To supplement mathematics for large models, submit the open source MathPile corpus with 9.5 billion tokens, which can also be used commercially

To supplement mathematics for large models, submit the open source MathPile corpus with 9.5 billion tokens, which can also be used commercially

Mar 01, 2024 pm 10:31 PM
industry mathpile

In the current development of intelligent dialogue models, powerful underlying models play a crucial role. The pre-training of these advanced models often relies on high-quality and diverse corpora, and how to build such a corpus has become a major challenge in the industry.

In the high-profile field of AI for Math, the relative scarcity of high-quality mathematical corpus limits the potential of generative artificial intelligence in mathematical applications.

#To address this challenge, the Generative Artificial Intelligence Laboratory of Shanghai Jiao Tong University launched “MathPile”. This is a high-quality, diverse pre-training corpus specifically targeted at the field of mathematics, which contains approximately 9.5 billion tokens and is designed to improve the capabilities of large models in mathematical reasoning.

In addition, the laboratory also launched the commercial version of MathPile - "MathPile_Commercial" to further broaden its application scope and commercial potential.

To supplement mathematics for large models, submit the open source MathPile corpus with 9.5 billion tokens, which can also be used commercially

  • Paper address: https://huggingface.co/papers/2312.17120

  • Project address: https://gair-nlp.github.io/MathPile/
  • Code address: https://github.com/GAIR-NLP/MathPile

Dataset address:

  • Research use: https://huggingface.co/datasets /GAIR/MathPile
  • Commercial version: https://huggingface.co/datasets/GAIR/MathPile_Commercial

Introduction

MathPile has the following characteristics:

To supplement mathematics for large models, submit the open source MathPile corpus with 9.5 billion tokens, which can also be used commercially

1. Mathematically centered: Different from the corpora that focused on general fields in the past, such as Pile, RedPajama, or the multilingual corpus ROOTS, etc., MathPile focuses on the field of mathematics. Although there are already some specialized mathematical corpora, they are either not open source (such as the corpus used by Google to train Minerva, OpenAI's MathMix), or are not rich and diverse enough (such as ProofPile and the recent OpenWebMath).

2. Diversity: MathPile has a wide range of data sources, such as open source mathematics textbooks, class notes, synthetic textbooks, and mathematics on arXiv Related papers, math-related entries on Wikipedia, lemma proofs and definitions on ProofWiki, high-quality math questions and answers on StackExchange, and math web pages from Common Crawl. The above content covers content suitable for primary and secondary schools, universities, graduate students, and mathematics competitions. MathPile covers 0.19B tokens of high-quality mathematics textbooks for the first time.

3. High Quality: The research team follows the concept of "less is more" during the collection process and firmly believes in data quality Better than quantity, even in the pre-training phase. From a data source of ~520B tokens (approximately 2.2TB), they went through a rigorous and complex set of pre-processing, pre-filtering, language identification, cleaning, filtering and deduplication steps to ensure the high quality of the corpus. It is worth mentioning that the MathMix used by OpenAI only has 1.5B tokens.

4. Data documentation: In order to increase transparency, the research team documented MathPile and provided a dataset sheet. During the data processing process, the research team also performed "quality annotation" on documents from the Web. For example, the score of language recognition and the ratio of symbols to words in the document allow researchers to further filter documents according to their own needs. They also performed downstream test set contamination detection on the corpus to eliminate samples from benchmark test sets such as MATH and MMLU-STEM. At the same time, the research team also discovered that there are also a large number of downstream test samples in OpenWebMath, which shows that extra care should be taken when producing pre-training corpus to avoid invalid downstream evaluation.

To supplement mathematics for large models, submit the open source MathPile corpus with 9.5 billion tokens, which can also be used commercially

MathPile’s data collection and processing process.

Data processing details

Today, as competition in the field of large models becomes increasingly fierce, many technology companies no longer disclose their data; of data sources, matching, not to mention detailed preprocessing details. On the contrary, MathPile summarizes a set of data processing methods suitable for the Math field based on previous explorations.

In the data cleaning and filtering part, the specific steps adopted by the research team are:

  • Detect lines containing "lorem ipsum". If "lorem ipsum" is replaced by less than 5 characters in the line, remove the line;
  • Detect lines containing "javescript" ” and contains a line with "enable", "disable" or "browser", and the number of characters in the line is less than 200 characters, then filter out the line;
  • Filter out less than 200 characters. 10 words and lines containing "Login", "sign-in", "read more...", or "items in cart";
  • Filter out the proportion of uppercase words More than 40% of the documents;
  • Filter out the documents whose lines ending with an ellipsis account for more than 30% of the entire document;
  • Filter out the non- Documents where the proportion of letter words exceeds 80%;
  • Filter out documents whose average English word character length is outside the range of (3, 10);
  • Filter out documents that do not contain at least two stop words (such as the, be, to, of, and, that, have, etc.);
  • Filter out ellipses and Documents with a word proportion exceeding 50%;
  • Filter out documents in which the line starting with a bullet accounts for more than 90%;
  • Filter out documents with lines starting with bullets Documents of less than 200 characters after excluding spaces and punctuation;
  • ...

More processing details can be found in the paper.

In addition, the research team also provided many data samples during the cleaning process. The picture below shows the nearly duplicate documents in Common Crawl detected by the MinHash LSH algorithm (shown in pink highlights).

To supplement mathematics for large models, submit the open source MathPile corpus with 9.5 billion tokens, which can also be used commercially

As shown in the figure below, the research team discovered problems from the MATH test set (as highlighted in yellow) during the data leakage detection process.

To supplement mathematics for large models, submit the open source MathPile corpus with 9.5 billion tokens, which can also be used commercially

Data set statistics and examples

The following table shows the various components of MathPile According to the statistical information, you can find arXiv papers. Textbooks usually have longer document lengths, while documents on wikis are relatively shorter.

To supplement mathematics for large models, submit the open source MathPile corpus with 9.5 billion tokens, which can also be used commercially

The picture below is a sample document of a textbook in the MathPile corpus. It can be seen that the document structure is relatively clear and the quality is high.

To supplement mathematics for large models, submit the open source MathPile corpus with 9.5 billion tokens, which can also be used commercially

Experimental results

The research team also disclosed some preliminary experimental results. They performed further pre-training based on the currently popular Mistral-7B model. Then, it was evaluated on some common mathematical reasoning benchmark data sets through a few-shot prompting method. The preliminary experimental data that have been obtained so far are as follows:

To supplement mathematics for large models, submit the open source MathPile corpus with 9.5 billion tokens, which can also be used commercially

These test benchmarks cover all levels of mathematical knowledge, including primary school mathematics (such as GSM8K, TAL-SCQ5K-EN and MMLU- Math), high school mathematics (such as MATH, SAT-Math, MMLU-Math, AQuA, and MathQA), and college mathematics (such as MMLU-Math). Preliminary experimental results announced by the research team show that by continuing to pre-train on textbooks and Wikipedia subsets in MathPile, the language model has achieved considerable improvements in mathematical reasoning capabilities at different difficulty levels.

The research team also emphasized that relevant experiments are still ongoing.

Conclusion

MathPile has received widespread attention since its release and has been reproduced by many parties. It is currently listed on the Huggingface Datasets trend list. The research team stated that they will continue to optimize and upgrade the data set to further improve data quality.

To supplement mathematics for large models, submit the open source MathPile corpus with 9.5 billion tokens, which can also be used commercially

MathPile is trending on Huggingface Datasets. To supplement mathematics for large models, submit the open source MathPile corpus with 9.5 billion tokens, which can also be used commercially
MathPile was forwarded by the well-known AI blogger AK, source: https://twitter.com/_akhaliq/status/1740571256234057798.

Currently, MathPile has been updated to the second version, aiming to contribute to the research and development of the open source community. At the same time, its commercial version of the data set has also been opened to the public.

The above is the detailed content of To supplement mathematics for large models, submit the open source MathPile corpus with 9.5 billion tokens, which can also be used commercially. For more information, please follow other related articles on the PHP Chinese website!

Statement of this Website
The content of this article is voluntarily contributed by netizens, and the copyright belongs to the original author. This site does not assume corresponding legal responsibility. If you find any content suspected of plagiarism or infringement, please contact admin@php.cn

Hot AI Tools

Undresser.AI Undress

Undresser.AI Undress

AI-powered app for creating realistic nude photos

AI Clothes Remover

AI Clothes Remover

Online AI tool for removing clothes from photos.

Undress AI Tool

Undress AI Tool

Undress images for free

Clothoff.io

Clothoff.io

AI clothes remover

Video Face Swap

Video Face Swap

Swap faces in any video effortlessly with our completely free AI face swap tool!

Hot Tools

Notepad++7.3.1

Notepad++7.3.1

Easy-to-use and free code editor

SublimeText3 Chinese version

SublimeText3 Chinese version

Chinese version, very easy to use

Zend Studio 13.0.1

Zend Studio 13.0.1

Powerful PHP integrated development environment

Dreamweaver CS6

Dreamweaver CS6

Visual web development tools

SublimeText3 Mac version

SublimeText3 Mac version

God-level code editing software (SublimeText3)

Hot Topics

Java Tutorial
1655
14
PHP Tutorial
1253
29
C# Tutorial
1227
24
DeepMind robot plays table tennis, and its forehand and backhand slip into the air, completely defeating human beginners DeepMind robot plays table tennis, and its forehand and backhand slip into the air, completely defeating human beginners Aug 09, 2024 pm 04:01 PM

But maybe he can’t defeat the old man in the park? The Paris Olympic Games are in full swing, and table tennis has attracted much attention. At the same time, robots have also made new breakthroughs in playing table tennis. Just now, DeepMind proposed the first learning robot agent that can reach the level of human amateur players in competitive table tennis. Paper address: https://arxiv.org/pdf/2408.03906 How good is the DeepMind robot at playing table tennis? Probably on par with human amateur players: both forehand and backhand: the opponent uses a variety of playing styles, and the robot can also withstand: receiving serves with different spins: However, the intensity of the game does not seem to be as intense as the old man in the park. For robots, table tennis

The first mechanical claw! Yuanluobao appeared at the 2024 World Robot Conference and released the first chess robot that can enter the home The first mechanical claw! Yuanluobao appeared at the 2024 World Robot Conference and released the first chess robot that can enter the home Aug 21, 2024 pm 07:33 PM

On August 21, the 2024 World Robot Conference was grandly held in Beijing. SenseTime's home robot brand "Yuanluobot SenseRobot" has unveiled its entire family of products, and recently released the Yuanluobot AI chess-playing robot - Chess Professional Edition (hereinafter referred to as "Yuanluobot SenseRobot"), becoming the world's first A chess robot for the home. As the third chess-playing robot product of Yuanluobo, the new Guoxiang robot has undergone a large number of special technical upgrades and innovations in AI and engineering machinery. For the first time, it has realized the ability to pick up three-dimensional chess pieces through mechanical claws on a home robot, and perform human-machine Functions such as chess playing, everyone playing chess, notation review, etc.

Claude has become lazy too! Netizen: Learn to give yourself a holiday Claude has become lazy too! Netizen: Learn to give yourself a holiday Sep 02, 2024 pm 01:56 PM

The start of school is about to begin, and it’s not just the students who are about to start the new semester who should take care of themselves, but also the large AI models. Some time ago, Reddit was filled with netizens complaining that Claude was getting lazy. "Its level has dropped a lot, it often pauses, and even the output becomes very short. In the first week of release, it could translate a full 4-page document at once, but now it can't even output half a page!" https:// www.reddit.com/r/ClaudeAI/comments/1by8rw8/something_just_feels_wrong_with_claude_in_the/ in a post titled "Totally disappointed with Claude", full of

At the World Robot Conference, this domestic robot carrying 'the hope of future elderly care' was surrounded At the World Robot Conference, this domestic robot carrying 'the hope of future elderly care' was surrounded Aug 22, 2024 pm 10:35 PM

At the World Robot Conference being held in Beijing, the display of humanoid robots has become the absolute focus of the scene. At the Stardust Intelligent booth, the AI ​​robot assistant S1 performed three major performances of dulcimer, martial arts, and calligraphy in one exhibition area, capable of both literary and martial arts. , attracted a large number of professional audiences and media. The elegant playing on the elastic strings allows the S1 to demonstrate fine operation and absolute control with speed, strength and precision. CCTV News conducted a special report on the imitation learning and intelligent control behind "Calligraphy". Company founder Lai Jie explained that behind the silky movements, the hardware side pursues the best force control and the most human-like body indicators (speed, load) etc.), but on the AI ​​side, the real movement data of people is collected, allowing the robot to become stronger when it encounters a strong situation and learn to evolve quickly. And agile

ACL 2024 Awards Announced: One of the Best Papers on Oracle Deciphering by HuaTech, GloVe Time Test Award ACL 2024 Awards Announced: One of the Best Papers on Oracle Deciphering by HuaTech, GloVe Time Test Award Aug 15, 2024 pm 04:37 PM

At this ACL conference, contributors have gained a lot. The six-day ACL2024 is being held in Bangkok, Thailand. ACL is the top international conference in the field of computational linguistics and natural language processing. It is organized by the International Association for Computational Linguistics and is held annually. ACL has always ranked first in academic influence in the field of NLP, and it is also a CCF-A recommended conference. This year's ACL conference is the 62nd and has received more than 400 cutting-edge works in the field of NLP. Yesterday afternoon, the conference announced the best paper and other awards. This time, there are 7 Best Paper Awards (two unpublished), 1 Best Theme Paper Award, and 35 Outstanding Paper Awards. The conference also awarded 3 Resource Paper Awards (ResourceAward) and Social Impact Award (

Li Feifei's team proposed ReKep to give robots spatial intelligence and integrate GPT-4o Li Feifei's team proposed ReKep to give robots spatial intelligence and integrate GPT-4o Sep 03, 2024 pm 05:18 PM

Deep integration of vision and robot learning. When two robot hands work together smoothly to fold clothes, pour tea, and pack shoes, coupled with the 1X humanoid robot NEO that has been making headlines recently, you may have a feeling: we seem to be entering the age of robots. In fact, these silky movements are the product of advanced robotic technology + exquisite frame design + multi-modal large models. We know that useful robots often require complex and exquisite interactions with the environment, and the environment can be represented as constraints in the spatial and temporal domains. For example, if you want a robot to pour tea, the robot first needs to grasp the handle of the teapot and keep it upright without spilling the tea, then move it smoothly until the mouth of the pot is aligned with the mouth of the cup, and then tilt the teapot at a certain angle. . this

Distributed Artificial Intelligence Conference DAI 2024 Call for Papers: Agent Day, Richard Sutton, the father of reinforcement learning, will attend! Yan Shuicheng, Sergey Levine and DeepMind scientists will give keynote speeches Distributed Artificial Intelligence Conference DAI 2024 Call for Papers: Agent Day, Richard Sutton, the father of reinforcement learning, will attend! Yan Shuicheng, Sergey Levine and DeepMind scientists will give keynote speeches Aug 22, 2024 pm 08:02 PM

Conference Introduction With the rapid development of science and technology, artificial intelligence has become an important force in promoting social progress. In this era, we are fortunate to witness and participate in the innovation and application of Distributed Artificial Intelligence (DAI). Distributed artificial intelligence is an important branch of the field of artificial intelligence, which has attracted more and more attention in recent years. Agents based on large language models (LLM) have suddenly emerged. By combining the powerful language understanding and generation capabilities of large models, they have shown great potential in natural language interaction, knowledge reasoning, task planning, etc. AIAgent is taking over the big language model and has become a hot topic in the current AI circle. Au

Hongmeng Smart Travel S9 and full-scenario new product launch conference, a number of blockbuster new products were released together Hongmeng Smart Travel S9 and full-scenario new product launch conference, a number of blockbuster new products were released together Aug 08, 2024 am 07:02 AM

This afternoon, Hongmeng Zhixing officially welcomed new brands and new cars. On August 6, Huawei held the Hongmeng Smart Xingxing S9 and Huawei full-scenario new product launch conference, bringing the panoramic smart flagship sedan Xiangjie S9, the new M7Pro and Huawei novaFlip, MatePad Pro 12.2 inches, the new MatePad Air, Huawei Bisheng With many new all-scenario smart products including the laser printer X1 series, FreeBuds6i, WATCHFIT3 and smart screen S5Pro, from smart travel, smart office to smart wear, Huawei continues to build a full-scenario smart ecosystem to bring consumers a smart experience of the Internet of Everything. Hongmeng Zhixing: In-depth empowerment to promote the upgrading of the smart car industry Huawei joins hands with Chinese automotive industry partners to provide

See all articles